Bad AI: Secrets and Lies in a Silicon Valley Corporation
Pros
None for IC's. They are scapegoats and vulnerable. For managers, it is an ideal place as they are totally free of any responsibilities. Rest and Vest.
Cons
Even before the recent trend of AI, the teams already used AI-related code names (X-AI, like AAI, BAI, CAI) for their models, even though most of them are based on linear models. The performance (predictive power) is poor, and the brand names of "AI" are very misleading. What AI means to Adobe is not high-performing ML models, but brand names! Now, in the era of real AI, the company is no inventor of AI. It uses the AI's invented by other companies, fine tune the models for the company's own use case, for instance firefly, and claim it is "a world leading AI company". While ironically, the model performance is no better than their open source counterpart, such as standard diffusion. The discussion between firefly and other Gen AI of image is very interesting. 1)When talking about the low performance of firefly, the company argues that firefly is trained on Adobe Stock only and is "ethical". 2)When the creator/contributor of the images of Adobe Stock is angry about firefly using their image to train the model, the company argues that since the training set is so big, it is hard to prove that firefly violates the copy rights of the creators. But if 2) holds true, the same logic should also be applied to other Gen AI's and 1) becomes questionable again. Adobe is never a [big data] company, (it is a big company, it is a data company, but it cannot handle big data!) the company can hardly handle training data with millions of rows, unless it uses very simple models, like linear models, or drop a significant amount of data. In fact one of the reason why firefly only uses images in Adobe Stock is that the team cannot handle large data sets (but the company uses moral issue as an excuse and even a selling point!) Adobe's models are bad, they have very low predictive power; Adobe's AI's are bad, their performance is infinitely close to free/open source AI's. Whatever this company does is impression management. No matter how bad their AI products are, they just need to give the financial market/Wall Street an impression that it is developing AI, then the financial analysts (who know nothing about AI) will flatter the company, and the price of ADBE will hit another historical high. Mission complete. Glassdoor removed my review last year, and I can imagine it will do the same to this review.
